OpenAI to put invisible watermark on ChatGPT text in the EU, Malaysians won’t get it by default

KUALA LUMPUR (Oct 12) — When two university students use ChatGPT to write the same assignment, one student in Petaling Jaya and the other in Munich, only one will have an invisible watermark. OpenAI plans to add the watermark to ChatGPT and Codex text in the EU only, as part of the EU AI Act's transparency rules. The watermark, called textGrain, subtly alters the model's word choices while still allowing a special detector to pick up on the pattern.

OpenAI will roll out the watermark over the coming weeks to ChatGPT and Codex users on all plans in the EU, but it will not be a default feature at launch. Malaysian developers can opt in to watermarking for select models from today, even though the feature is off by default. OpenAI is also working with cloud partners to offer watermarking for model outputs accessed through their services in the coming weeks.

While the watermark does not identify the user, it is not foolproof. OpenAI's own tests show that it can be weakened, with detection rates dropping significantly when words are replaced with synonyms. The watermark is also easier to weaken in short passages, mathematics answers, and translated text. OpenAI will initially grant detector access only to approved researchers and expert organizations, not to students or teachers.

The watermark can also indicate that an OpenAI system generated or processed part of a passage, but it does not tell you about the level of human creativity or judgement. OpenAI is not the first AI company to act on provenance, as Google, Meta, Microsoft, and Anthropic have also implemented similar measures for their respective AI tools.